Road extraction in off-road environments poses significant challenges due to uneven terrain, unstructured class boundaries, irregular features, and strong textures. While deep learning methods for free-space detection have been developed over the past decade, most focus on urban settings rather than the more complex off-road scenarios. Current segmentation methods struggle to detect blurred road boundaries in these environments. To address these challenges, this paper introduces OFF-CSUNet, a novel network featuring an improved transformer block that alternates between cross-shaped and sliding window self-attention, expanding the interactive field of local self-attention. OFF-CSUNet integrates this enhanced transformer block into a U-shaped network architecture, utilizing an encoder-decoder structure and skip connections to effectively gather local and global information. Additionally, a cross-attention fusion module is implemented to dynamically combine RGB images and LiDAR point cloud data for accurate off-road road detection. Experiments conducted on the ORFD dataset demonstrate that our model outperforms existing methods for off-road free-space detection.